AI system

On-device

Two small models live on the bracelet's M55:

1. Feature extractor — converts raw sensor windows (1 minute) into a feature vector.
2. Activity classifier — labels the minute (sleep, walking, running, sedentary, exercise).

Both are int8 quantised, ~80 KB each.

On-phone

The phone runs the personalisation engine:

1. Sleep stager — segments the night into wake/light/deep/REM.
2. WVI computer — applies the four sub-scores against the member's baseline.
3. Direction picker — ranks the five archetypes and picks one.

These run nightly on-device after sync. Member's raw data never leaves the phone unless cloud backup is opted in.

In the cloud (optional)

If the member opts in to cloud backup:

1. Daily summaries are encrypted with the member's key and stored.
2. Aggregated cohort statistics (anonymised, k-anonymity ≥ 50) are used to retrain the global priors.

That is the entire cloud ML footprint. There is no global per-user model in the cloud.

What we do not do

• We do not run an LLM "coach" that reads your raw data.
• We do not stream sensor data to the cloud in real time.
• We do not use member data to train models that benefit other members without explicit consent.